Claude proved Fermat's Last Theorem in Lean, and the person who was already doing it says it adds nothing mathematically

Two facts sit next to each other and neither cancels the other out.

Anthropic published on September 4 that an internal general-purpose research model, roughly comparable to Claude Fable 5.1, formalized Fermat's Last Theorem in Lean over 11 days working largely autonomously. The artifact: 13 million lines of Lean, 30,300 proved theorems, 29,500 of them used in the final proof, about 6 billion output tokens of compute. That codebase is around 5x the size of Mathlib, the community formalization library that took a decade of human effort to build (Anthropic).

Kevin Buzzard runs the EPSRC-funded FLT formalization project at Imperial College. He has been working on this for years. On September 4 he posted that he pulled the codebase, compiled it, ran comparator against it, and it checks out. He also ran security checks, having an agent inspect the non-mathematical code and personally reviewing anything that looked odd. His verdict: it proves the theorem with no assumptions beyond the axioms of mathematics (Xena Project).

Then he says it adds essentially nothing mathematically. The proof uses the 1995 Darmon-Diamond-Taylor approach, not modern methods. It covers exponents p≥17, which is sufficient because the smaller irregular primes were already formalized. And it faithfully transcribes existing literature. Nobody learned a new theorem. What happened is that a very large amount of extremely tedious, extremely precise transcription got done by something that doesn't get bored.

I've been reading his post trying to figure out which fact I care about more, and I think it's the second one, but not for the reason people are using it. "It adds nothing mathematically" isn't a dismissal. It's a description of exactly the work I want an agent doing. Transcription with zero tolerance for error, at a scale where humans stop being able to hold it in their heads, verified by a compiler that doesn't care how confident anyone feels.

The detail almost nobody is covering: three personal Claude Max plans, driving a platform called Prove2Me, formalized Vinogradov's Three Primes Theorem in three days. Three consumer subscriptions. Three days. That number changes the story from datacenter stunt to workflow, and it's the one I'd act on.

The pattern to copy is the verification loop, not the math. Buzzard didn't trust the announcement, he compiled it. The Lean kernel is what makes that possible, and it's the reason this result is checkable at all while every other "AI did science" claim this year required you to take someone's word for it. When you're evaluating agent output at scale, the question is whether you have a Lean kernel for your domain. Type checkers, property tests, SPICE simulations, deterministic build-and-run. If you don't have one, you're grading vibes and the volume will bury you.

For contrast on what these models still can't do: Epoch AI launched FrontierMath Erdős on September 1, built from 68 unsolved Erdős problems with a $300 and 72-hour budget per attempt. GPT-6 Astra scored 3% by disproving one problem and proving another. GPT-5.6 Sol, GPT-5.5, Fable 5.1 and Fable 5 all ran out of budget with no verified proof and scored zero (Epoch AI). Transcription at superhuman volume, yes. New mathematics, barely.